Complex language
Keywords and fixed SOPs handle standard questions but struggle with vague needs, variants, and follow-ups.
Back to selected projects
Enterprise AI product · PRD plan
AI Sales Agent
An intelligent customer-service and sales-assistance product for independent e-commerce sites
LLMs handle natural-language understanding, RAG grounds answers in merchant knowledge, workflows control key processes, and business APIs connect SKU, order, and customer data for consultation, clarification, recommendation, and order service.
AI SALES AGENT / PRD01
Business problem
Moving beyond standard Q&A requires more than replacing the model
Independent stores have distinct products, policies, and brand language. AI needs merchant-specific knowledge, dynamic data, conversation continuity, profile updates, and reliable fallback to enter the sales-service flow.
Merchant-specific knowledge
SKU, product facts, policies, and brand language cannot be answered from one generic knowledge set.
Service-sales gap
Passive answers do not uncover budget, recipient, use case, and preferences needed for recommendations.
Disconnected data
Consultation signals are not structured into profiles, while knowledge, orders, and customers remain siloed.
02
Controlled workflow
Make every response follow a traceable business path
The agent does not improvise freely. Intent, retrieval, dynamic queries, and output validation form an explicit route. Hover or focus a node to inspect its responsibility and output.
Agent / AIRules and systemsHuman fallback
- 01Message intakeRules and systemsReceive and preprocess user input
Normalized query and required context
- 02Intent detectionAgent / AIIdentify consultation, recommendation, order, after-sales, or human intent
Business intent and confidence
- 03Workflow routingRules and systemsChoose RAG, API, clarification, or SOP based on rules
Controlled processing path
- 04Knowledge and dataRules and systemsRetrieve stable knowledge and query live business data
Grounded context and business results
- 05Prompt + LLMAgent / AICombine profile, context, and constraints
Natural-language response candidate
- 06Output validationRules and systemsCheck evidence, rules, and exception states
Returnable answer or fallback signal
- 07Human handoffHuman fallbackHandle low confidence, complaints, or repeated failure
Reliable service continuation
03
Product decisions
Place generation inside business boundaries
Separate stable knowledge and live data
RAG manages descriptions, FAQs, and policies; APIs provide inventory, orders, logistics, and some prices.
High-volatility facts cannot rely on static knowledge alone.Control critical nodes with workflows
Intent, rules, and conditions route retrieval, queries, replies, and human handoff.
Refunds, complaints, and order exceptions do not rely solely on model judgment.Minimize profile inference
Store only explicit or high-confidence budget, preference, scenario, and SKU signals.
Model guesses cannot become confirmed profile fields.Degrade safely
When RAG, APIs, or models fail, fall back through keywords, FAQ, Rule/SOP, and human service.
Never invent product facts without reliable evidence.04
AI capability architecture
Load merchant knowledge, rules, and data on a shared AI foundation
The architecture lets each merchant configure its own knowledge base, prompts, workflows, and brand voice while connecting its SKU, customer, and order systems.
Understanding
Intent, references, requirement extraction, and clarification.
Orchestration
Coze Workflow, business rules, SOPs, and handoff conditions.
Knowledge and data
Private RAG, vector retrieval, SKU, order, logistics, and customer APIs.
Generation and validation
Prompt assembly, LLM response, fact constraints, and fallback.
- 01User input
- 02Intent / rules
- 03Workflow
- 04RAG / API / profile
- 05Prompt
- 06LLM
- 07Validation
Structured PRD
Product requirements document
This web version is organized from the existing Sales Expert Agent PRD. It preserves positioning, boundaries, data strategy, fallback, and test standards without inventing implementation results or production metrics.Positioning and core value
AI Sales Agent is an intelligent customer-service and sales-assistance product for independent e-commerce merchants. LLMs handle language, RAG supplies grounded merchant knowledge, workflows constrain service paths, and profiles plus business data support consultation, recommendation, orders, and sales assistance.
Value for both sides
Merchant
Reduce repetitive support work, unify knowledge and rules, and configure knowledge, prompts, workflows, and brand voice per merchant.
Customer
Avoid repeating needs and receive continuous consultation and recommendations based on budget, preferences, and use cases.
Target users and scenarios
Store customer
Browse, ask about products, receive recommendations, order, track, and access after-sales support.
Independent merchant
Manage product knowledge, service configuration, brand rules, customer operations, and business-data connections.
Core scenarios
- Product facts
- Need clarification
- Personalized recommendation
- Multi-turn comparison
- Price and inventory
- Order and logistics
- After-sales FAQ
- Human handoff
Core capability modules
ModulePrimary ownerResponsibility
Intent detectionLLM + rules
Identify consultation, recommendation, logistics, after-sales, and human-support intent.
Coze WorkflowProcess system
Control retrieval, data queries, response generation, and fallback.
RAG knowledgeKnowledge system
Provide private product, FAQ, brand, and after-sales knowledge.
User profileProfile service
Store confirmed budget, preference, use case, watched SKU, and concerns.
Business APIBusiness system
Fetch SKU, inventory, order, logistics, and customer data.
Prompt + LLMAgent
Assemble role, rules, context, knowledge, and business data into a response.
Human handoffSupport agent
Handle unreliable, high-risk, or explicitly requested human cases.
RAG processing chain
Knowledge quality must be diagnosed along the retrieval chain. Without reliable retrieval, the system must not generate unverified product facts.
- 1Data
- 2Chunking
- 3Embedding
- 4Vector Retrieval
- 5Rerank
- 6Context
- 7Prompt
- 8LLM
Conversation context and profile
The session retains context for references such as “this one,” “the second item,” or “something cheaper.” Profiles update only from explicit or high-confidence information.
- Budget
- Category
- Color
- Style
- Material
- Recipient
- Use case
- Watched SKU
- Core concern
Data freshness layers
Choose access methods by update frequency so live business facts do not become stale knowledge.
Descriptions, FAQ, policies, brand information
Synchronization + RAGInventory, orders, logistics, some prices
Business API or real-time syncException handling and fallback
When retrieval, APIs, or models cannot finish reliably, the system states the issue and enters a controlled next path.
Unknown intentRequest more information or a problem category
No valid RAG resultDo not invent; clarify or hand off
API failureSuggest retry or human support
Knowledge conflictPrefer the business system or verified source
Repeated failureTransfer to human support
Test and acceptance criteria
These are PRD-defined validation directions and test cases, not claimed production accuracy or conversion results.
Product fact accuracy
- Retrieve the matching SKU
- Match the knowledge source
- Do not mix product facts
Multi-turn context
- Resolve follow-up references
- Keep the current SKU context
- Understand comparisons
Vague recommendation
- Ask for missing requirements
- Stay within budget
- Use current merchant SKUs
Missing knowledge
- Do not generate unverified facts
- State that data is unavailable
- Support clarification or human confirmation
Profile use
- Reuse confirmed budget
- Prefer the latest explicit statement
- Never store inference as fact
Product summary